John Langford leaves Microsoft after AI Frontiers folds into Microsoft AI

The researcher who founded Microsoft Research's AI Frontiers group says the group is over. John Langford is a senior principal researcher at MSR. He is also a long-time figure in machine learning. He posted on hunch.net on 6 November that he has left Microsoft. He says the rest of the founding AI Frontiers leadership team has reached the same decision.

AI Frontiers was set up inside MSR in 2023, after ChatGPT, Langford writes. Its job was organised research into agents and new model capabilities. The group was about 50 people. He says they worked on the Autogen agent orchestration framework and the Fara computer-use agent. They also worked on new ideas such as the Next Latent training objective and Dion3, an improved Muon optimiser. Another was the Full Bandwidth Transformer, which uses recursion to add power without adding parameters.

The reorg that ended it

In early September, AI Frontiers was folded into Microsoft AI, Langford writes. Microsoft AI is the consumer AI group that ex-Inflection co-founder Mustafa Suleyman has run since 2024. Its leadership is set on scaling as the way to improve models, Langford says. He says it is "aggressively not research".

Returns to scale are more predictable than returns to research, Langford grants. But he argues the maths still favours research. An example he cites: a variant that halves training and inference cost for the same capability. The industry's AI spend is near one trillion US dollars, he writes. Even a small cost cut at that spend is a large number.

AI Frontiers also tried to share with the rest of MSR. It built a monorepo that was shared by default with every Microsoft Research team, Langford writes. That monorepo held a pretraining codebase he had talked about in a COLM keynote the month before. He says at least four other model-building pipelines are being written inside Microsoft, three of them in private repos. One shared codebase should have reduced that work.

Why he did not return to MSR

Langford says he spoke to people about going back into Microsoft Research itself. He did not take that path. Two reasons, in his words. He could not get the AI Frontiers work used, or seriously considered, inside Microsoft AI. He does not expect a new MSR team to fare better. The second reason: the basic MSR deal, do good work and good things happen, looks broken to him. His team was reorganised against its will.

He closes by thanking the people he worked with. The year ahead will have no AI Frontiers interns, he says. The post reached 81 points on Hacker News within a few hours. Readers there asked who else had left and whether the Autogen and Fara projects will keep moving.

Source

The End of AI Frontiers, John Langford, hunch.net. Hacker News discussion.